Wang, L. (2024) - Mammography with Deep Learning for Breast Cancer Detection

Summary

• Comprehensive review of medical imaging and deep learning (DL) applications for Breast Cancer Detection.
• Evaluates performance across public datasets (DDSM, CBIS-DDSM, INbreast, MIAS, BreaKHis).

Notes

Medical Imaging Techniques

• Mammography
  • Utilises low-dose X-rays to generate breast images that radiologists use to find lumps, calcifications and distortions.
  • Recommended for women over 40.
  • Limitations:
    • reduced sensitivity for women with dense breast tissue.
• Digital Breast Tomosynthesis (DBT)
  • Uses x-rays to generate 3d breast images, useful for detecting BC in dense breasts.
  • Faces inter-observer variability which can affect accuracy.
• Ultrasound Imaging
  • High-frequency sound waves to produce detailed images of breast abnormalities, especially in women with dense breasts.
  • Does not involve radiation, so it's generally safer.
• Magnetic Resonance Imaging (MRI)
  • Uses magnetic fields and radio waves to image soft tissue without ionizing radiation.
  • Effective for dense breasts and high-risk patients.
  • Tends to have lower Specificity (more false positives)
• Positron Emission Tomography (PET)
  • Detects gamma rays emitted by radiotracers to map metabolic activity and identify malignant lesions.
  • Creates 3d images of internal body.

Deep learning techniques

• DL architectures covered:
  • CNN
  • Transfer Learning
  • Ensemble Learning
    • Stacking
    • Boosting
    • Bagging
  • Attention based models (Transformer)
    • SE-Net
    • Channel Attention Networks (CAN)

Preprocessing

• Denoising techniques:
  • Median filter
  • Wiener Filter
  • Non-local means filter
  • Total variation (TV) denoising
  • Wavelet-based denoising
  • Gaussian filter
  • Anisotropic Diffusion
  • BM3D Denoising
  • Auto-Encoder
• Normalisation methods typically applied after denoising, like min-max normalisation

Datasets

• DDSM – Original digitised mammography dataset.
• Lee et al. (2017) - A curated mammography data set for use in computer-aided detection and diagnosis research – Curated subset with improved annotations.
• INbreast, MIAS, and BreaKHis – Alternative datasets with varied imaging quality and pathology balance.
• Data augmentation (rotations, flips, noise injection) helps overcome limited sample sizes.

Performance Metrics

• Accuracy
  • measures the proportion of correct predictions made by the model.
  • $\frac{TP + TN}{TP+TN+FP+FN}$
• Precision
  • Measures the proportion of true positive predictions out of all positive predictions made by the model.
  • $\frac{TP}{TP + FP}$
• Sensitivity
  • Measures the proportion of true positive predictions out of all actual positive cases in the dataset.
  • $\frac{TP}{TP + FN}$
• F1-Score
  • A composite metric that balances Precision and Sensitivity.
  • $2 \times \frac{Precision \times Recall}{Precision + Recall}$
• Area Under the Curve (AUC)
  • Evaluates the model’s ability to distinguish between positive and negative cases across various threshold values.
  • A higher AUC indicates better discriminatory performance.
• Mean-Squared Error
  • Measures the average squared difference between predicted and actual values, used primarily for regression tasks.
  • $\frac{1}{n}\sum_{i=1}^{n}(y_i - \hat{y}_i)^2$

Key Findings

• Deep learning significantly improves sensitivity and specificity compared to conventional CAD systems.
• Transfer learning and ensembles increase performance across heterogeneous datasets.
• Attention-based and Transformer architectures further enhance lesion localisation and reduce false positives.
• Model interpretability is essential for radiologist trust and clinical adoption.

Limitations and Future Directions

• Performance still constrained by dataset bias and limited public data availability.
• Need for cross-institutional generalisation and federated learning approaches.
• Integration of explainability, uncertainty estimation, and radiologist feedback loops recommended for deployment.

Related

• Lee et al. (2017) - A curated mammography data set for use in computer-aided detection and diagnosis research – describes the CBIS-DDSM dataset used in multiple models reviewed.
• Chollet, 2018 – reference for CNN design and training workflow.

Reference

• Wang, L. (2024). Mammography with deep learning for breast cancer detection. Frontiers in Oncology, 14, 1281922. https://doi.org/10.3389/fonc.2024.1281922